Orchard scene-oriented multi-robot fruit tree task allocation problem optimization method

Abstract

The invention discloses a multi-robot fruit tree task allocation problem optimization method for an orchard scene, and the method comprises the steps: S1, numbering task points of each mature fruit tree, and recording the coordinates and yield of the task points; s2, starting to record the running time; s3, carrying out initialization by using a load-distance balance initialization mechanism; s4, calculating an energy consumption value corresponding to each solution; s5, executing an experience adaptive selection strategy; s6, implementing a clustering-based local search mechanism on the solution to be optimized; and S7, repeatedly executing the steps S4-S6 until a termination condition is reached. According to the invention, through a load-distance balance initialization method, the quality of an initial solution is improved, and the calculation overhead of subsequent optimization is reduced; a clustering local search mechanism based on K-means is adopted, and a task distribution structure in a route is effectively identified and optimized; an experience adaptive selection strategy is introduced, the selection probability is dynamically adjusted according to historical optimization experience, and the convergence efficiency of the algorithm is improved; and the task allocation mechanism based on the population is used for allocating the obtained multiple trips to the robot.

Description

Technical Field

[0001] The present invention belongs to the technical field of path planning, and in particular relates to an optimization method for multi-robot fruit tree task allocation problems in orchard scenarios. Background Art

[0002] In recent years, the booming development of smart agricultural technologies is reshaping traditional agricultural production models. As a major agricultural country, my country is actively promoting agricultural modernization, leveraging its vast arable land resources to implement intelligent transformation. Henan Province, for example, has an agricultural development plan that aims to reach 15 million acres of high-quality orchards by 2025. This goal demonstrates my country's unwavering confidence in the development of modern agriculture.

[0003] However, in the process of agricultural modernization, labor shortages have become a key bottleneck hindering development. This is particularly true in the cash crop sector, where the production of high-value fruits such as apples, pears, and peaches still relies heavily on manual labor. This problem has become increasingly prominent with the acceleration of urbanization—the large-scale migration of rural labor to cities has led to a severe labor shortage in agricultural production. At the same time, changes in the global economic situation have driven labor costs upwards, placing unprecedented economic pressure on the agricultural sector.

[0004] Throughout the fruit cultivation process, harvesting accounts for the largest portion of manual labor. To address this challenge, researchers are turning their attention to the application of automation technologies. The development of agricultural robots, in particular, offers new possibilities for addressing the challenges of fruit harvesting. Research teams both domestically and internationally have developed a variety of fruit harvesting robot prototypes and have achieved positive results in experimental settings. These robots are capable of autonomously completing complex tasks such as fruit identification, sorting, and harvesting, opening up new avenues for reducing reliance on human labor.

[0005] As single-robot technology matures, the concept of multi-robot collaborative systems has emerged. By deploying multiple robots to operate simultaneously, these systems significantly improve work efficiency while also effectively overcoming the performance bottlenecks of single robots when handling large-scale tasks. Multi-robot systems achieve parallel operations through task allocation and collaborative mechanisms, significantly improving overall operational efficiency.

[0006] However, current multi-robot systems still have significant shortcomings in task allocation. Traditional allocation strategies are often oversimplified, primarily basing tasks on physical distance or basic classification. This crude allocation approach ignores the inherent characteristics and complexity of tasks and fails to effectively address workload variations across tasks. For example, allocating tasks based solely on distance can easily lead to uneven allocation of system resources, causing some robots to operate at excessive capacity while others remain idle.

[0007] Therefore, optimizing the task allocation mechanism for multi-robot systems is key to improving harvesting efficiency and reducing operating costs. This requires establishing a more comprehensive evaluation system that takes into account multiple factors, such as task time consumption, energy requirements, and deadlines. By thoroughly analyzing task characteristics and integrating them with the robots' real-time status and operational capabilities, and building an intelligent task scheduling mechanism, we can truly achieve optimal allocation of system resources and improve overall operational efficiency. Summary of the Invention

[0008] In order to solve the above problems, the present invention proposes an optimization method for multi-robot fruit tree task allocation problem in orchard scenarios, which combines an adaptive, experience-based discrete genetic algorithm to improve the efficiency of collaborative work of multiple robots while reducing energy consumption costs.

[0009] The technical solution adopted in the present invention is:

[0010] An optimization method for multi-robot fruit tree task allocation problem in an orchard scenario includes the following steps:

[0011] S1, number each mature fruit tree task point and record its coordinates and yield;

[0012] S2, start recording the running time;

[0013] S3, initialized using the load-distance balancing initialization mechanism;

[0014] S4, calculate the energy consumption value corresponding to each solution according to the following formula:

[0015]

[0016] Among them, Z represents the total energy consumption, L i Indicates the load of the robot after completing the i-th task. For x ij and b i is defined as follows:

[0017]

[0018] S5, executing the experience-adaptive selection strategy;

[0019] S6, implements a clustering-based local search mechanism for the optimization solution;

[0020] S7, repeat steps S4-S6 until the termination condition is reached.

[0021] Furthermore, the specific steps of S3 are:

[0022] S31, input the basic parameters required by the mechanism, including the task node set N = {1, 2, ..., n} ∪ {0}, where 0 represents the robot's starting position; the population size Psize; the distance matrix between tasks d ij ; Output of each task q i ;Robot carrying capacity Q and robot weight W;

[0023] S32, according to the distance matrix d ij Sort the tasks in descending order according to their distance from the warehouse to obtain the sequence S1 and the index matrix Rank1 corresponding to each task point in the sequence;

[0024] S33, according to the task output q i Sort in descending order to obtain the sequence S2 and the index matrix Rank2 corresponding to each task point in the sequence;

[0025] S34, execute for each solution p in the population:

[0026] S341, calculate the dynamic weight coefficient: β p =(p-1) / (Psize-1), where p∈(1,Psize);

[0027] S342, calculate the comprehensive ranking value for each task i: Rank i =β p ×Rank1(i)+(1-β p )×Rank2(i);

[0028] S343, according to Rank i Sorting in ascending order generates a comprehensive task sequence SC;

[0029] S344, construct feasible solutions;

[0030] S35, repeat step S34 until all solutions in the population are initialized.

[0031] Furthermore, in step S344, the feasible solution construction includes the following steps:

[0032] S3441, initialize the required parameters, including the task route set Current route and the current load L = 0;

[0033] S3442, perform task assignment on each solution in turn:

[0034] S34421, select the first task i from the comprehensive task sequence SC;

[0035] S34422, determine the load constraint: if Li+q i+1≤Q, then execute S3423, otherwise the robot returns to the warehouse to unload;

[0036] S34423, update current route: Li = Li + q i+1 ;CR=CR∪{i};Remove task i from the comprehensive task sequence SC;

[0037] S34424, constructing the feasible task set FT;

[0038] S34425, select the best task: obtain the task k with the smallest weighted order in the feasible task set FT; if the distance d between task k and the currently executed task is kt Less than the threshold d th , then add task k to the current route CR, where the threshold d th is the distance from the current task to the warehouse; otherwise, the robot returns to the warehouse to unload;

[0039] S3443, repeat step S3442 until all tasks in the comprehensive task sequence SC of the current solution are assigned and a total of Route paths are obtained, so that p=p+1.

[0040] Furthermore, in step S34424, the steps of constructing the feasible task set FT are:

[0041] S344241, Initialize the feasible task set

[0042] S344242, for each task j in the remaining tasks of the comprehensive task sequence SC, if L+q j ≤Q, then add each task j to the feasible task set FT;

[0043] S344243, calculate task occupancy rate Pr = q j / Q, j∈FT;

[0044] S344244, sort the tasks in the feasible task set FT according to the calculated task occupancy rate Pr.

[0045] Furthermore, the specific steps of S5 are:

[0046] S51, during the initial optimization, directly select the solution bestInd corresponding to the minimum energy consumption value as the current solution, and record the energy consumption value Z(bestInd) of the current optimal solution;

[0047] S52, initializing the optimization probability of each solution;

[0048] S53, performing adaptive selection;

[0049] S54, evaluate the solution and update the experience matrix.

[0050] Further, the step of initializing the optimization probability of each solution in S52 is as follows:

[0051] S521, generate a selection range ranges = {0, 10%,..., SR}, where 10% means randomly selecting one from the solutions ranked in the top 10% of the energy consumption values in the population for subsequent optimization. According to experience, SR is set to 60%, that is, the solution to be optimized is randomly selected from the solutions ranked in the top 60% at most;

[0052] S522, initialize the successful experience record vector archive = {0, 0,..., 0};

[0053] S523, construct a normalized probability distribution archiveF:

[0054]

[0055] S524, calculate the selection probability corresponding to each selection range:

[0056]

[0057] where c = 1 / Psize.

[0058] Further, the specific steps of performing adaptive selection in S53 are as follows: [[ID=2']]

[0059] S531, select the selection range ranges(i) with the largest probability according to the F value;

[0060] S532, calculate the sampling position: k = ranges(i) × Psize;

[0061] S533, select the top k optimal individuals from the population Pop to form an elite set ES.

[0062] Further, the specific steps of evaluating the solution and updating the experience matrix in S54 are as follows:

[0063] S541, randomly select a solution Ind from the elite set ES; !

[0064] S542, perform step S6 on Ind to obtain newInd;

[0065] S543, calculate the energy consumption value Z(newInd) of newInd according to step S4;

[0066] S544, if Z(newInd) < Z(bestInd), then update the experience record: archive(i) = archive(i) + 1, update the optimal solution: bestInd = newInd;

[0067] S545 , recalculate the normalized probability distribution archiveF according to step S523 .

[0068] Furthermore, the specific steps of S6 are:

[0069] S61, set the parameter MaxDistance to 0;

[0070] S62, evaluate each path r in all routes of the solution in turn:

[0071] S621, use the K-means algorithm to divide the tasks in route r into two clusters A and B;

[0072] S622, calculate the cluster center coordinates:

[0073] S6221, C1=(∑x i / n1,∑y i / n1), where (x i ,y i ) is the coordinate of the i-th task in cluster A, n1 is the number of tasks in cluster A;

[0074] S6222, C2=(∑x j / n2,∑y j / n2), where (x j ,y j ) is the coordinate of the jth task in cluster B, n2 is the number of tasks in cluster B;

[0075] S623, calculate the inter-cluster distance D(r): D(r) = || C1 - C2 ||;

[0076] S624, if D(r) is greater than the current maximum distance MaxDistance, then update MaxDistance = D(r); set r as the target optimized route TR;

[0077] S63, repeat S62 until all paths in Route are evaluated;

[0078] S64, optimizing the target optimization route TR using the 2-opt method.

[0079] The beneficial effects produced by the present invention are:

[0080] 1. Improve the quality of the initial solution and reduce the computational overhead of subsequent optimization through the load-distance balance initialization method;

[0081] 2. Use a K-means-based clustering local search mechanism to effectively identify and optimize the task distribution structure in the route;

[0082] 3. Introducing an experience-adaptive selection strategy to dynamically adjust the selection probability based on historical optimization experience to improve the convergence efficiency of the algorithm;

[0083] 4. A population-based task allocation mechanism is used to assign the obtained multiple trips to robots. BRIEF DESCRIPTION OF THE DRAWINGS

[0084] Figure 1 is a flow chart of the method of the present invention;

[0085] Figure 2 Schematic diagram of 2-opt in step S64 of the present invention. DETAILED DESCRIPTION

[0086] The present invention will be further described below with reference to the accompanying drawings.

[0087] like Figure 1 As shown in FIG, a multi-robot fruit tree task allocation problem optimization method for an orchard scenario includes the following steps:

[0088] S1, number each mature fruit tree task point and record its coordinates and yield;

[0089] S2, start recording the running time;

[0090] S3, initialized using the load-distance balancing initialization mechanism;

[0091] S31, input the basic parameters required by the mechanism, including the task node set N = {1, 2, ..., n} ∪ {0}, where 0 represents the robot's starting position, such as the warehouse; the population size Psize; the distance matrix between tasks d ij ; Output of each task q i ;Robot carrying capacity Q and robot weight W;

[0092] S32, according to the distance matrix d ij Sort the tasks in descending order according to their distance from the warehouse to obtain the sequence S1 and the index matrix Rank1 corresponding to each task point in the sequence;

[0093] S33, according to the task output q i Sort in descending order to obtain the sequence S2 and the index matrix Rank2 corresponding to each task point in the sequence;

[0094] S34, execute for each solution p in the population:

[0095] S341, calculate the dynamic weight coefficient: β p =(p-1) / (Psize-1), where p∈(1,Psize);

[0096] S342, calculate the comprehensive ranking value for each task i: Rank i =β p ×Rank1(i)+(1-β p )×Rank2(i);

[0097] S343, according to Rank i Sorting in ascending order generates a comprehensive task sequence SC;

[0098] S344, construct a feasible solution. The specific steps are as follows:

[0099] S3441, initialize the required parameters, including the task route set Current route and the current load L = 0;

[0100] S3442, perform task assignment on each solution in turn:

[0101] S34421, select the first task i from the comprehensive task sequence SC;

[0102] S34422, determine the load constraint: if Li+q i+1 ≤Q, then execute S3423, otherwise the robot returns to the warehouse to unload;

[0103] S34423, update current route: Li = Li + q i+1 ;CR=CR∪{i};Remove task i from the comprehensive task sequence SC;

[0104] S34424, construct the feasible task set FT, the specific steps are:

[0105] S344241, Initialize the feasible task set

[0106] S344242, for each task j in the remaining tasks of the comprehensive task sequence SC, if L+q j ≤Q, then add each task j to the feasible task set FT;

[0107] S344243, calculate task occupancy rate Pr = q j / Q, j∈FT;

[0108] S344244, sorting the tasks in the feasible task set FT according to the calculated task occupancy rate Pr;

[0109] S34425, select the best task: obtain the task k with the smallest weighted order in the feasible task set FT; if the distance d between task k and the currently executed task is kt Less than the threshold d th, then add task k to the current route CR, where the threshold d th is the distance from the current task to the warehouse; otherwise, the robot returns to the warehouse to unload;

[0110] S3443, repeat step S3442 until all tasks in the comprehensive task sequence SC of the current solution are assigned and a total of Route paths are obtained, so that p=p+1;

[0111] S35, repeat step S34 until all solutions in the population are initialized.

[0112] S4, calculate the energy consumption value corresponding to each solution according to the following formula:

[0113]

[0114] Among them, Z represents the total energy consumption, L i Indicates the load of the robot after completing the i-th task. For x ij and b i is defined as follows:

[0115]

[0116] S5, executing the experience-adaptive selection strategy;

[0117] S51, during the initial optimization, directly select the solution bestInd corresponding to the minimum energy consumption value as the current solution, and record the energy consumption value Z(bestInd) of the current optimal solution;

[0118] S52, initializing the optimization probability of each solution;

[0119] S521: Generate a selection range ranges = {0, 10%, ..., SR}, where 10% means randomly selecting a solution from the top 10% of the energy consumption values in the population for subsequent optimization. Based on experience, SR is set to 60%, meaning that the solution to be optimized is randomly selected from the top 60% of the solutions at most. Furthermore, the selection range is initialized only during the first run of the mechanism, and subsequent selection ranges remain consistent with the initial one.

[0120] S522, initialize the success experience record vector archive = {0, 0, ..., 0}; in addition, it is initialized only during the first generation of operation, and the records updated in step S544 are used subsequently;

[0121] S523, construct normalized probability distribution archiveF:

[0122]

[0123] S524. Calculate the selection probability corresponding to each selection range:

[0124]

[0125] where c = 1 / Psize.

[0126] S53. Perform adaptive selection;

[0127] S531. Select the selection range ranges(i) with the highest probability according to the F value;

[0128] S532. Calculate the sampling position: k = ranges(i) × Psize;

[0129] S533. Select the top k optimal individuals from the population Pop to form the elite set ES.

[0130] S54. Evaluate the solution and update the experience matrix.

[0131] S541. Randomly select a solution Ind from the elite set ES;

[0132] S542. Perform step S6 on Ind to obtain newInd;

[0133] S543. Calculate the energy consumption value Z(newInd) of newInd according to step S4;

[0134] S544. If Z(newInd) < Z(bestInd), update the experience record: archive(i) = archive(i) + 1, and update the optimal solution: bestInd = newInd;

[0135] S545. Recalculate the normalized probability distribution archiveF according to step S523.

[0136] S6. Implement a clustering-based local search mechanism for the solution to be optimized;

[0137] S61. Set the parameter MaxDistance to 0;

[0138] S62. Evaluate each path r in all Routes of the solution in turn:

[0139] S621. Use the K-means algorithm to divide the tasks in route r into two clusters A and B; <​​​​​​​ / n1), where (x i ,y i ) is the coordinate of the i-th task in cluster A, n1 is the number of tasks in cluster A;

[0142] S6222, C2=(∑x j / n2,∑y j / n2), where (x j ,y j ) is the coordinate of the jth task in cluster B, n2 is the number of tasks in cluster B;

[0143] S623, calculate the inter-cluster distance D(r): D(r) = || C1 - C2 ||;

[0144] S624, if D(r) is greater than the current maximum distance MaxDistance, then update MaxDistance = D(r); and set r as the target optimized route TR;

[0145] S63, repeat S62 until all paths in Route are evaluated;

[0146] S64, using the 2-opt method to optimize the target optimization route TR, such as Figure 2 As shown;

[0147] S7, repeat steps S4-S6 until the termination condition is reached.

[0148] This paper proposes a dynamic weighted task sorting method that comprehensively considers load and distance; designs a local search mechanism based on task clustering to improve the efficiency of route optimization; and innovatively introduces an empirical adaptive selection strategy to achieve dynamic adjustment of selection pressure.

[0149] To further illustrate the superiority of the present invention in solving multi-objective multi-robot task allocation, Table 1 shows the results obtained by the present invention and some excellent multi-robot task allocation algorithms, Multi-objective Discrete ArtificialBee Colony (MODABC), Non-dominated Sorting Genetic Algorithm-Ⅱ (NSGA-Ⅱ) and Multi-objective Evolutionary Algorithm based on Decomposition (MOEA / D), on the generated orchard multi-robot task allocation test set.

[0150] Table 1 Comparison of the optimal times achieved by experimental results of datasets

[0151]

[0152] The examples provide experimental results on a generated test set of orchard multi-objective, multi-robot task allocation problems. These problems vary in difficulty, including greenhouses ranging in size from 50×50 to 70×70 square meters, containing 625, 900, and 1225 fruit trees, respectively. The fruit tree maturity rate in each scenario was set to 0.8. Therefore, each combination of the number of fruit trees and their maturity represented a different test problem to evaluate the algorithm's task allocation performance under different work scenarios and task difficulties. The number of evaluations in the experiment was set to 10,000, and the population size was set to 10. To avoid the influence of randomness on the experimental results, each algorithm was run 10 times for each test problem. The evaluation results were metrically evaluated using hypervolume. Table 1 shows the number of problems for which each algorithm achieved optimal results in each test problem scenario. By comparison, it can be seen that the method proposed in the present invention demonstrates superior optimization performance in solving multi-objective, multi-robot task allocation. Specifically, regardless of the test problem scenario, the number of optimal results obtained was greater than the number of optimal results obtained by other algorithms.

[0153] In summary, the present invention can effectively handle the multi-objective multi-robot task allocation problem by combining the random neighborhood search method with the genetic algorithm, and provide decision makers with a series of ideal compromise solutions.

[0154] A number of multi-objective evolutionary algorithms have been designed and widely recognized for their effectiveness in solving multi-objective problems. However, to solve more complex problems, the structure of multi-objective optimization algorithms also becomes complex. Therefore, it is necessary to improve the performance of algorithms without significantly increasing their complexity to broaden their application. Because single-objective optimization algorithms are less complex than multi-objective ones, a hybrid framework combining single-objective and multi-objective optimization algorithms can achieve better performance with lower complexity.

[0155] This paper combines the variable neighborhood search theory with a population-based approach, and proposes a random neighborhood search method combined with a genetic algorithm to solve the multi-objective multi-robot task allocation problem in an orchard scenario. Experiments and analysis demonstrate the effectiveness and superiority of the algorithm. The problem-solving ability of the algorithm is influenced by a variety of methods and mechanisms, including a clustering-based task initialization method, a task sequence optimization method, a task balancing mechanism, a neighboring task optimization mechanism, and a population-based task allocation mechanism. Due to the high correlation between operator design and specific scenarios, this method cannot be directly applied to other task allocation problems. Nevertheless, the design process, innovative mechanisms, and problem-solving methods of the algorithm provide valuable insights for the development of other task allocation algorithms.

[0156] Although the above describes the specific embodiments of the present invention in conjunction with the accompanying drawings, it is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art on the basis of the technical solution of the present invention without any creative work are still within the scope of protection of the present invention.

Claims

1. An optimization method for multi-robot fruit tree task allocation problem in orchard scenarios, characterized by: The following steps are involved: S1, number each mature fruit tree task point and record its coordinates and yield; S2, start recording the running time; S3, initialized using the load-distance balancing initialization mechanism; S4, calculate the energy consumption value corresponding to each solution according to the following formula: Among them, Z represents the total energy consumption, L i Indicates the load of the robot after completing the i-th task. For x ij and b i is defined as follows: S5, executing the experience-adaptive selection strategy; S6, implements a clustering-based local search mechanism for the optimization solution; S7, repeat steps S4-S6 until the termination condition is reached.

2. The method for optimizing the multi-robot fruit tree task allocation problem in an orchard scenario according to claim 1 is characterized in that: The specific steps of S3 are: S31, input the basic parameters required by the mechanism, including the task node set N = {1, 2, ..., n} ∪ {0}, where 0 represents the robot's starting position; the population size Psize; the distance matrix between tasks d ij ; Output of each task q i ;Robot carrying capacity Q and robot weight W; S32, according to the distance matrix d ij Sort the tasks in descending order according to their distance from the warehouse to obtain the sequence S1 and the index matrix Rank1 corresponding to each task point in the sequence; S33, according to the task output q i Sort in descending order to obtain the sequence S2 and the index matrix Rank2 corresponding to each task point in the sequence; S34, execute for each solution p in the population: S341, calculate the dynamic weight coefficient: β p =(p-1) / (Psize-1), where p∈(1,Psize); S342, calculate the comprehensive ranking value for each task i: Rank i =β p ×Rank1(i)+(1-β p )×Rank2(i); S343, according to Rank i Sorting in ascending order generates a comprehensive task sequence SC; S344, construct feasible solutions; S35, repeat step S34 until all solutions in the population are initialized.

3. The method for optimizing the multi-robot fruit tree task allocation problem in an orchard scenario according to claim 2 is characterized in that: In step S344, the feasible solution construction includes the following steps: S3441, initialize the required parameters, including the task route set Current route and the current load L = 0; S3442, perform task assignment on each solution in turn: S34421, select the first task i from the comprehensive task sequence SC; S34422, determine the load constraint: if Li+q i+1 ≤Q, then execute S3423, otherwise the robot returns to the warehouse to unload; S34423, update current route: Li = Li + q i+1 ;CR=CR∪{i};Remove task i from the comprehensive task sequence SC; S34424, constructing the feasible task set FT; S34425, select the best task: obtain the task k with the smallest weighted order in the feasible task set FT; if the distance d between task k and the currently executed task is kt Less than the threshold d th , then add task k to the current route CR, where the threshold d th is the distance from the current task to the warehouse; otherwise, the robot returns to the warehouse to unload; S3443, repeat step S3442 until all tasks in the comprehensive task sequence SC of the current solution are assigned and a total of Route paths are obtained, so that p=p+1.

4. The method for optimizing the multi-robot fruit tree task allocation problem in an orchard scenario according to claim 3 is characterized in that: In step S34424, the steps of constructing the feasible task set FT are: S344241, Initialize the feasible task set S344242, for each task j in the remaining tasks of the comprehensive task sequence SC, if L+q j ≤Q, then add each task j to the feasible task set FT; S344243, calculate task occupancy rate Pr = q j / Q, j∈FT; S344244, sort the tasks in the feasible task set FT according to the calculated task occupancy rate Pr.

5. The method for optimizing the multi-robot fruit tree task allocation problem in an orchard scenario according to claim 1 is characterized in that: The specific steps of S5 are: S51, during the initial optimization, directly select the solution bestInd corresponding to the minimum energy consumption value as the current solution, and record the energy consumption value Z(bestInd) of the current optimal solution; S52, initializing the optimization probability of each solution; S53, performing adaptive selection; S54, evaluate the solution and update the experience matrix.

6. The orchard-oriented multi-robot fruit tree task allocation problem optimization method according to claim 5 is characterized in that: The step of initializing the optimization probability of each solution in S52 is: S521, generating a selection range ranges = {0, 10%, ..., SR}, where 10% means randomly selecting a solution from the top 10% of the energy consumption values in the population for subsequent optimization; S522, initialize the success experience record vector archive = {0, 0, ..., 0}; S523, construct normalized probability distribution archiveF: S524, calculate the selection probability corresponding to each selection range: Where c = 1 / Psize.

7. The orchard-oriented multi-robot fruit tree task allocation problem optimization method according to claim 5 is characterized in that: The specific steps of performing the adaptive selection in S53 are: S531, select the selection range ranges(i) with the highest probability according to the F value; S532, calculate the sampling position: k = ranges(i) × Psize; S533, select the top k best individuals from the population Pop to form the elite set ES.

8. The orchard-oriented multi-robot fruit tree task allocation problem optimization method according to claim 5 is characterized in that: The specific steps of evaluating the solution and updating the experience matrix in S54 are: S541, randomly select a solution Ind from the elite set ES; S542, execute step S6 on Ind to obtain newInd; S543, calculating the energy consumption value Z(newInd) of newInd according to step S4; S544, if Z(newInd) < Z(bestInd), update the experience record: archive(i) = archive(i) + 1, and update the optimal solution: bestInd = newInd; S545, recalculate the normalized probability distribution archiveF according to step S523.

9. The orchard-based multi-robot fruit tree task allocation optimization method according to claim 1 is characterized in that: The specific steps of S6 are as follows: S61, set the parameter MaxDistance to 0; S62, evaluate each path r in all Routes of the solution in sequence: S621, use the K-means algorithm to divide the tasks in route r into two clusters A and B; S622, calculate the cluster center coordinates: S6221, C1=(∑x i / n1,∑y i / n1), where (x i ,y i ) is the coordinate of the i-th task in cluster A, n1 is the number of tasks in cluster A; S6222, C2=(∑x j / n2,∑y j / n2), where (x j ,y j ) is the coordinate of the jth task in cluster B, n2 is the number of tasks in cluster B; S623, calculate the inter-cluster distance D(r): D(r) = ||C1 - C2||; S624, if D(r) is greater than the current maximum distance MaxDistance, update MaxDistance = D(r); set r as the target optimization route TR; S63, repeat S62 until all paths in Route have been evaluated; S64, optimize the target optimization route TR using the 2-opt method.

Images